Why it matters

Streaming quantiles power all latency monitoring. Exact percentiles impossible at scale.

Advertisement

The architecture

GK for eps-approx.

t-digest for cluster-based.

DDSketch for relative error.

Streaming quantilesGKeps-approxt-digestcluster-basedDDSketchrelative errorDDSketch (Datadog) mergeable and relative-error; t-digest common in metrics
Streaming quantiles.
Advertisement

How it works end to end

GK sketch.

t-digest.

DDSketch.